OpenAI is out here trying to kill my GrepGoblin AI detection startup: https://t.co/0fHzFjvvou
(Interesting insight into the internal process of model personality shaping.)
My new co-authored paper offers 3 evidence-based tips for lawyer AI use:
1. Use AI only when you can assess, explain, & build on output
2. Limit AI use to narrow, well-defined parts of project
3. Limit AI use under tight deadlines or cognitive fatigue. https://t.co/7X4OKp7uMW
@OrinKerr@ARozenshtein@IThinkIAgree This study sheds some light, showing that even MFA-trained writers preferred AI-written excerpts when the AI was fine-tuned on the works of a famous author. When the tell-tale voice of the assistant disappears, so too does some of the aesthetic distaste.
https://t.co/XWcCmm1zXY
I'm grateful to @lsolum for engaging with and promoting my work on AI and law.
I've made an audio version of the paper available here: https://t.co/xd0EV52eNx— and a critical faculty workshop made with my app enTalkenator here: https://t.co/JhrIGbFcvl
https://t.co/IyVFUxezHR
Learning the technical details of the transformer and other deep learning architectures and then building a big app that leverages them were critical to my developing anything substantial to say about AI. I've learned so much from so many (especially @jackclarkSF and @karpathy).
This paper results from experiences that go all the way back to my math days, when I (embarassingly in retrospect but consistent with the sometimes-smugness of a young math grad student) scoffed at curve-fitting phenomena as a substitute for understanding them. 2/x
My new paper on modern AI. I argue: (1) It thinks, and denying this is dangerous. (2) Models are individual alien intelligences. (3) They deceptively activate the intentional stance. (4) We need a socially-supported alien stance to meet this challenge. (1/2)
These are deeply fascinating papers. I added intro classes and an academic workshop on them to the enTalkenator podcast feed (generated with Sonnet 3.7 in enTalkenator): https://t.co/Ax4zGblD9f
New Anthropic research: Tracing the thoughts of a large language model.
We built a "microscope" to inspect what happens inside AI models and use it to understand Claude’s (often complex and surprising) internal mechanisms.
I guess it’s just my luck that Anthropic is having API issues the day after a lot of new people find out about the app… Claude Sonnet 3.7 is still the best model to use with enTalkenator, but if it hasn’t been working for you this morning, it’s because it’s down.
"When Artificial Intelligence Workshops Your Article For You—And It's Really Useful" My first post at a new blog about a new app. cc:@christor
https://t.co/P0ya2n2xYG
@an_average_bear@OrinKerr I’ve had a paper (many papers actually!) on the back burner for awhile. Put my brief outline/idea into the app, selected the First Draft template and got 20,000 words out that some law review editors would’ve gone for after adding footnotes.
@an_average_bear@OrinKerr Concerns? Yes. Objections? Depends on how. The app includes built in templates that generate original papers (10-20k words), response papers, and brutal critical responses. But you can modify these. Upshot is that you can put in an outline and get a “normal science” paper out.
"When Artificial Intelligence Workshops Your Article For You—And It's Really Useful" My first post at a new blog about a new app. cc:@christor
https://t.co/P0ya2n2xYG
@ProfFerguson@OrinKerr Funny you ask. In some versions of the workshop template, I told some participants they might have more a comment than a question. But even using Claude Sonnet (which is what I always select to do these), they said “maybe this is more a comment than a question” annoyingly often.
This week is @OrinKerr's fascinating article on data scanning and 4A, following a midweek intro class on a groundbreaking study relating to dark matter. (You should hear/read the response papers it generates to things like a train ticket...) Recommend listening at at least 1.5x.
I've built an iOS app, enTalkenator, initially intended just to generate audio narrations from PDFs of academic articles. It now, from any text, generates academic workshops, new papers, podcasts, debates, and introductory classes. I describe it here: https://t.co/gh7Z6NLWTG